The Autonomous Network
Roadmap, not current topology. Production uses one hosted Core and a custodial worker payout rail. The sections below describe staged design directions; deployed contract primitives do not make the workflow autonomous.
Every AI API you’ve used can be unplugged. Stripe can freeze your account. AWS can suspend your inference endpoint. OpenAI can deprecate a model out from under your product. That’s not a flaw in their operations — it’s the shape of centralized services.
AIPG is working toward a generation layer without a permanent central coordinator. Today, workers authenticate to one hosted core, jobs route through Redis/WebSockets, den is recorded off-chain, and a treasury sender pays workers AIPG on Base. Deployed contracts and assignment-bound validator evidence provide building blocks for a more distributed system; they do not yet remove the core or its operational trust.
What’s Already On-Chain
These implementations are deployed on Base today as facets of the Grid Diamond contract. Deployment is not the same as an active production workflow:
WorkerRegistry— registry implementation; production workers currently authenticate in core and the new bonding facet still has a gated deploy pathJobAnchor— anchoring implementation; no production batch publisher currently anchors every jobModelVault— deployed model records and reference hashes; live model availability still comes from connected workersRecipeVault— workflow registry primitive; current core primarily loads reviewed local recipes unless chain sync is configuredRewardPool,DenReporter,PaymentRouter— deployed claim primitives; the live payout rail is still custodial AIPG transfersStakingVault— deployed passive-staking vault; that rewards program has endedRoleManager— on-chain contract roles; core still has its own authenticated application roles and keys
Look them up in Developers for live addresses.
What’s Coming
Validator-attested job routing
Today the API node picks which worker handles a job. One proposed later stage lets workers self-claim jobs through a deterministic mechanism (described in Claim Resolution) while validators attest to objective evidence. That design is not implemented as the production router.
Smart-contract payments
A future audited settlement design could move generation fees through a contract instead of the current off-chain service. Worker and validator fee rules, claim timing, and dispute behavior are not finalized.
Bonded validators
The big trust upgrade. Validators start as an evidence-only audit layer, then move toward stake-backed assignments once targeted probes, reference checks, quorum, and dispute tooling exist. See Validator Nodes.
Dispute resolution
When attestations disagree, a future dispute protocol could use bounded re-execution and objective evidence. No production fraud-proof contract, automatic loser slashing, or deterministic no-human dispute path is live.
Why On-Chain Routing Matters
Most “decentralized AI” projects are decentralized in their compute layer and centralized everywhere else. There’s still an API node that decides who runs what, a payment processor in the loop, and a team that can — under pressure — yank the rug.
Putting routing, settlement, and verification into smart contracts changes the shape of the network:
| What | Before (centralized) | After (autonomous) |
|---|---|---|
| Job assignment | API node picks a worker | Workers self-claim via deterministic on-chain rule |
| Payment settlement | Off-chain payment processor | Smart-contract escrow + release |
| Result verification | ”Trust the worker” | Validator evidence, quorum, and eventual objective-fraud slashing |
| Disputes | Email the team | Fraud-proof in contract, deterministic outcome |
| Failure mode | Provider can turn it off | Network keeps running while workers exist |
The sequencing analogy is useful, but it is not proof of the destination. AIPG’s hosted Core is today’s coordinator. Replacing that trust requires partner-operated federation, deterministic state replication, validator independence, audited settlement, and observed failover evidence.
Confidential Computing as the Final Piece
Even if routing, settlement, and validator commitments move on-chain, one open question remains: who can see the prompts? On ordinary hardware, the worker operator and host OS can.
Confidential Computing is the proposed privacy lane: prompts and outputs would be decrypted only inside an attested GPU TEE. This is not the privacy guarantee of ordinary workers or current production requests.
The target architecture combines federated routing, on-chain economic commitments, independent validator evidence, and hardware-isolated privacy. Each property needs its own implementation and production evidence before the network can claim it.
Read the Deep Design
The full on-chain economics proposal — sampling rates, slashing math, dispute resolution, fee splits — lives at Future On-Chain Architecture. It is a design proposal, not the source of truth for currently deployed behavior. Current behavior lives in the Grid core, contract address inventory, and operator runbooks.
For day-to-day operator and developer docs while the autonomous layer rolls out, see P2P Mode and For Developers.